id	author	title	date	pages	extension	mime	words	sentences	flesch	summary	cache	txt
harper-generative-2020		harper-generative-2020			.docx	application/vnd.openxmlformats-officedocument.wordprocessingml.document	5838	489	59	Figure 2 Images generated with a simple statistical model appear as noise as the model is insufficient to capture the structure of the real data (Markov chains trained using wine bottles and circles from Google's QuickDraw dataset). Other types of generative statistical models, like Naive Bayes or a higher-order Markov chain,[footnoteRef:1] could perhaps capture a bit more information about the training data, but they would still be insufficient for real-world applications like this.[footnoteRef:2] Image, video, and audio are complicated; it is hard to reduce them to their essence with basic statistical rules in the way we were able to with the ordering of letters in English and Italian. Figure 4 A GAN being trained on wine bottle sketches from Google's quickdraw dataset (https://github.com/googlecreativelab/quickdraw-dataset) shows the generator learning how to produce better sketches over time. GANs in Action: Deep Learning with Generative Adversarial Networks.	./cache/harper-generative-2020.docx	./txt/harper-generative-2020.txt